Autonomy Talks - Benoit Landry: Differentiable Optimization in Nonlinear Optimal Control

Autonomy Talks - 21/12/2020

Speaker: Benoit Landry, Autonomous Systems Lab, Stanford University

Title: Differentiable Optimization in Nonlinear Optimal Control

Abstract: Numerical optimization is at the core of countless robotics applications. Almost every single class of mathematical program - quadratic, semidefinite, mixed integer - has found its way into planning and control algorithms. Recently, a certain class of optimization problems called bilevel optimization, where two mathematical programs are nested into one another, have gained renewed attention thanks to advances in automatic differen- tiation frameworks and machine learning research. These advances mostly rely on the ability to make the nested optimization problem efficiently differentiable. In this talk, I will give a brief overview of our work on bilevel optimization, and its applications to the challenges of modern robotics. I will demonstrate how novel solution methods can be designed to leverage advances in differentiation tools while retaining the gains of state of the art constrained nonlinear optimization solvers. I will also demonstrate how reformu- lation of particularly challenging problems of nonlinear optimal control such as planning through contact, adversarial learning and Lyapunov synthesis can all be efficiently tackled as bilevel optimization problems.